1 //===- Promotion.cpp - Implementation of linalg Promotion -----------------===// 2 // 3 // Copyright 2019 The MLIR Authors. 4 // 5 // Licensed under the Apache License, Version 2.0 (the "License"); 6 // you may not use this file except in compliance with the License. 7 // You may obtain a copy of the License at 8 // 9 // http://www.apache.org/licenses/LICENSE-2.0 10 // 11 // Unless required by applicable law or agreed to in writing, software 12 // distributed under the License is distributed on an "AS IS" BASIS, 13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 14 // See the License for the specific language governing permissions and 15 // limitations under the License. 16 // ============================================================================= 17 // 18 // This file implements the linalg dialect Promotion pass. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #include "mlir/Dialect/Linalg/IR/LinalgOps.h" 23 #include "mlir/Dialect/Linalg/IR/LinalgTypes.h" 24 #include "mlir/Dialect/Linalg/Passes.h" 25 #include "mlir/Dialect/Linalg/Utils/Intrinsics.h" 26 #include "mlir/Dialect/Linalg/Utils/Utils.h" 27 #include "mlir/Dialect/LoopOps/LoopOps.h" 28 #include "mlir/EDSC/Helpers.h" 29 #include "mlir/IR/AffineExpr.h" 30 #include "mlir/IR/AffineExprVisitor.h" 31 #include "mlir/IR/AffineMap.h" 32 #include "mlir/IR/OpImplementation.h" 33 #include "mlir/Pass/Pass.h" 34 #include "mlir/Support/LLVM.h" 35 #include "mlir/Support/STLExtras.h" 36 #include "mlir/Transforms/FoldUtils.h" 37 38 #include "llvm/ADT/SetVector.h" 39 #include "llvm/Support/CommandLine.h" 40 41 using namespace mlir; 42 using namespace mlir::edsc; 43 using namespace mlir::edsc::intrinsics; 44 using namespace mlir::linalg; 45 using namespace mlir::linalg::intrinsics; 46 using namespace mlir::loop; 47 48 using llvm::SetVector; 49 50 #define DEBUG_TYPE "linalg-promotion" 51 52 static llvm::cl::OptionCategory clOptionsCategory(DEBUG_TYPE " options"); 53 static llvm::cl::opt<bool> clPromoteDynamic( 54 "test-linalg-promote-dynamic", 55 llvm::cl::desc("Test generation of dynamic promoted buffers"), 56 llvm::cl::cat(clOptionsCategory), llvm::cl::init(false)); 57 58 static AffineMap getAffineDifferenceMap(MLIRContext *context) { 59 AffineExpr d0(getAffineDimExpr(0, context)), d1(getAffineDimExpr(1, context)); 60 return AffineMap::get(2, 0, {d0 - d1}); 61 } 62 63 static Value *allocBuffer(Type elementType, Value *size, bool dynamicBuffers) { 64 auto *ctx = size->getContext(); 65 auto width = llvm::divideCeil(elementType.getIntOrFloatBitWidth(), 8); 66 if (!dynamicBuffers) 67 if (auto cst = dyn_cast_or_null<ConstantIndexOp>(size->getDefiningOp())) 68 return alloc( 69 MemRefType::get(width * cst.getValue(), IntegerType::get(8, ctx))); 70 Value *mul = muli(constant_index(width), size); 71 return alloc(MemRefType::get(-1, IntegerType::get(8, ctx)), mul); 72 } 73 74 // Performs promotion of a `subView` into a local buffer of the size of the 75 // *ranges* of the `subView`. This produces a buffer whose size may be bigger 76 // than the actual size of the `subView` at the boundaries. 77 // This is related to the full/partial tile problem. 78 // Returns a PromotionInfo containing a `buffer`, `fullLocalView` and 79 // `partialLocalView` such that: 80 // * `buffer` is always the size of the full tile. 81 // * `fullLocalView` is a dense contiguous view into that buffer. 82 // * `partialLocalView` is a dense non-contiguous slice of `fullLocalView` 83 // that corresponds to the size of `subView` and accounting for boundary 84 // effects. 85 // The point of the full tile buffer is that constant static tile sizes are 86 // folded and result in a buffer type with statically known size and alignment 87 // properties. 88 // To account for general boundary effects, padding must be performed on the 89 // boundary tiles. For now this is done with an unconditional `fill` op followed 90 // by a partial `copy` op. 91 static PromotionInfo promoteFullTileBuffer(OpBuilder &b, Location loc, 92 SubViewOp subView, 93 bool dynamicBuffers, 94 OperationFolder *folder) { 95 auto zero = constant_index(folder, 0); 96 auto one = constant_index(folder, 1); 97 98 auto viewType = subView.getType(); 99 auto rank = viewType.getRank(); 100 Value *allocSize = one; 101 SmallVector<Value *, 8> fullRanges, partialRanges; 102 fullRanges.reserve(rank); 103 partialRanges.reserve(rank); 104 for (auto en : llvm::enumerate(subView.getRanges())) { 105 auto rank = en.index(); 106 auto rangeValue = en.value(); 107 Value *d = rangeValue.size; 108 allocSize = muli(folder, allocSize, d).getValue(); 109 fullRanges.push_back(d); 110 partialRanges.push_back(range(folder, zero, dim(subView, rank), one)); 111 } 112 SmallVector<int64_t, 4> dynSizes(fullRanges.size(), -1); 113 auto *buffer = 114 allocBuffer(viewType.getElementType(), allocSize, dynamicBuffers); 115 auto fullLocalView = view( 116 MemRefType::get(dynSizes, viewType.getElementType()), buffer, fullRanges); 117 auto partialLocalView = slice(fullLocalView, partialRanges); 118 return PromotionInfo{buffer, fullLocalView, partialLocalView}; 119 } 120 121 SmallVector<PromotionInfo, 8> 122 mlir::linalg::promoteSubViews(OpBuilder &b, Location loc, 123 ArrayRef<Value *> subViews, bool dynamicBuffers, 124 OperationFolder *folder) { 125 if (subViews.empty()) 126 return {}; 127 128 ScopedContext scope(b, loc); 129 SmallVector<PromotionInfo, 8> res; 130 res.reserve(subViews.size()); 131 DenseMap<Value *, PromotionInfo> promotionInfoMap; 132 for (auto *v : subViews) { 133 SubViewOp subView = cast<SubViewOp>(v->getDefiningOp()); 134 auto viewType = subView.getType(); 135 // TODO(ntv): support more cases than just float. 136 if (!viewType.getElementType().isa<FloatType>()) 137 continue; 138 auto promotionInfo = 139 promoteFullTileBuffer(b, loc, subView, dynamicBuffers, folder); 140 promotionInfoMap.insert(std::make_pair(subView.getResult(), promotionInfo)); 141 res.push_back(promotionInfo); 142 } 143 144 for (auto *v : subViews) { 145 SubViewOp subView = cast<SubViewOp>(v->getDefiningOp()); 146 auto info = promotionInfoMap.find(v); 147 if (info == promotionInfoMap.end()) 148 continue; 149 // TODO(ntv): value to fill with should be related to the operation. 150 // For now, just use APFloat(0.0f). 151 auto t = subView.getType().getElementType().cast<FloatType>(); 152 Value *fillVal = constant_float(folder, APFloat(0.0f), t); 153 // TODO(ntv): fill is only necessary if `promotionInfo` has a full local 154 // view that is different from the partial local view and we are on the 155 // boundary. 156 fill(info->second.fullLocalView, fillVal); 157 } 158 159 for (auto *v : subViews) { 160 auto info = promotionInfoMap.find(v); 161 if (info == promotionInfoMap.end()) 162 continue; 163 copy(cast<SubViewOp>(v->getDefiningOp()), info->second.partialLocalView); 164 } 165 return res; 166 } 167 168 static void promoteSubViewOperands(LinalgOp op, SetVector<Value *> subViews, 169 bool dynamicBuffers, 170 OperationFolder *folder) { 171 // 1. Promote the specified views and use them in the new op. 172 OpBuilder b(op); 173 ScopedContext scope(b, op.getLoc()); 174 auto promotedBufferAndViews = promoteSubViews( 175 b, op.getLoc(), subViews.getArrayRef(), dynamicBuffers, folder); 176 SmallVector<Value *, 8> opViews; 177 opViews.reserve(op.getNumInputsAndOutputs()); 178 SmallVector<std::pair<Value *, Value *>, 8> writebackViews; 179 writebackViews.reserve(subViews.size()); 180 unsigned promotedIdx = 0; 181 for (auto *view : op.getInputsAndOutputs()) { 182 if (subViews.count(view) != 0) { 183 opViews.push_back(promotedBufferAndViews[promotedIdx].fullLocalView); 184 writebackViews.emplace_back(std::make_pair( 185 view, promotedBufferAndViews[promotedIdx].partialLocalView)); 186 promotedIdx++; 187 } else { 188 opViews.push_back(view); 189 } 190 } 191 192 // 2. Append all other operands as they appear, this enforces that such 193 // operands are not views. This is to support cases such as FillOp taking 194 // extra scalars etc. 195 auto operands = getAssumedNonViewOperands(op); 196 opViews.append(operands.begin(), operands.end()); 197 op.clone(b, op.getLoc(), opViews); 198 199 // 3. Emit write-back for the promoted output views: copy the partial view. 200 for (auto viewAndPartialLocalView : writebackViews) { 201 // Note: use the old op to determine whether the operand view is an output. 202 bool isOutput = 203 op.getIndexOfOutput(viewAndPartialLocalView.first).hasValue(); 204 if (isOutput) 205 copy(viewAndPartialLocalView.second, viewAndPartialLocalView.first); 206 } 207 208 // 4. Dealloc local buffers. 209 for (const auto &pi : promotedBufferAndViews) 210 dealloc(pi.buffer); 211 } 212 213 static void promoteSubViews(FuncOp f, bool dynamicBuffers) { 214 SmallVector<LinalgOp, 8> toErase; 215 OperationFolder folder(f.getContext()); 216 f.walk([dynamicBuffers, &folder, &toErase](LinalgOp op) { 217 // TODO(ntv) some heuristic here to decide what to promote. Atm it is all or 218 // nothing. 219 SetVector<Value *> subViews; 220 for (auto it : op.getInputsAndOutputs()) 221 if (auto sv = dyn_cast_or_null<SubViewOp>(it->getDefiningOp())) 222 subViews.insert(sv); 223 if (!subViews.empty()) { 224 promoteSubViewOperands(op, subViews, dynamicBuffers, &folder); 225 toErase.push_back(op); 226 } 227 }); 228 for (auto op : toErase) 229 op.erase(); 230 } 231 232 namespace { 233 struct LinalgPromotionPass : public FunctionPass<LinalgPromotionPass> { 234 LinalgPromotionPass() = default; 235 LinalgPromotionPass(bool dynamicBuffers) : dynamicBuffers(dynamicBuffers) {} 236 237 void runOnFunction() override { 238 promoteSubViews(getFunction(), dynamicBuffers); 239 } 240 241 bool dynamicBuffers; 242 }; 243 } // namespace 244 245 std::unique_ptr<OpPassBase<FuncOp>> 246 mlir::linalg::createLinalgPromotionPass(bool dynamicBuffers) { 247 return std::make_unique<LinalgPromotionPass>(dynamicBuffers); 248 } 249 250 static PassRegistration<LinalgPromotionPass> 251 pass("linalg-promote-subviews", "promote subview ops to local buffers", [] { 252 return std::make_unique<LinalgPromotionPass>(clPromoteDynamic); 253 }); 254